Reflection on Generative AI Poems 2012-2014

These writing samples are not strictly academic in nature but reflect over a decade of applied generative AI in creative processes. The following three poems used original python software developed during my undergraduate cognitive science studies at the University of California. The unpublished software, named “Words Masher,' implemented extremely rudimentary neural networks and markov chains trained on specific corpuses of literature sanitized and labeled by hand to produce text output. In each instance a single (most often randomly generated) word acted as a prompt and output length was set beforehand. The input was generated with a command that resembled: “x.generate_sentence("do", 24)”. The raw output was many pages of largely unformatted and probabilistically generated text used as creative inspiration, requiring extensive editing and license to produce cohesive poetry. Human editing acted as a sieve for novel imagery and required much more deletion than augmentation. The process depended heavily on “textual-pareidolia,” or creative pattern finding in the uninspired output. The pieces were produced between 2012 and 2013 and published in two printed Los Angeles art magazines in 2014 (with no AI attribution, an oversight). These poems are representative of a body of work published online and in print between 2010 and 2018.

Pseudepigrapha

The first sample was produced by a model trained on a corpus of five books, weighted most heavily on a translation of “Fragments” by Heraclitus and to a lesser extent on poetry by Philip Larkin and three other unrecorded texts. While even the final version requires some creative interpretation on behalf of the reader, it evokes some clearly human angst related to mortality, ignorance, and self-determination. It is regretful that I no longer have any examples of the raw output with which to compare the final draft. I do however have a small piece that was clearly edited but that did not make it into the final draft; I believe in retrospect the passage was omitted because it did not fit the pared down theme of the final poem and thus lacked lucidity. The omitted sample follows the poem below.

All things come unto a beardless lad, tripping: it is the end always beginning again. This unseen harmony is the law. Men act as if there were none, exclaiming "We are the Law!" This belief prevails as dreams do in sleep, every grown man a liar. Few know day from darkness, but death is wisdom’s shifted night, and Nature provides for all.

It awakens our awareness. Generations flow like blood; the limit of all sacrifice ends in the sea of thought. So deep are the waters and yet we bathe in mud. Knowing this is the greatest difficulty. It is all we are to expose it. One day we may learn from the few and good; but corpses are taught in periods of a thousand years. 

This world consists of Good and it's opposite in tension, like beasts. We treat this war as if it were a season but living this way is sickness. The king knows this, and will not overstep his fate. Man is a prince who must learn to live as a king. This truth is more than ears witness; our senses estrange us. 

False thoughts’ draught drips desire and pursues us into wine washed slumber. The man who knows this finds one thing above all: we are that which we think we are. The prize of the gods is hunger and eternity an unexpected feast; man decides his portions and pays with raving lips.

But even fish can be parched. The people must admit, therefore, that they are bad and asleep; weariness the same for young and old. Do not expect the whole width of time; demand to seek fresh water now.

The sun has heard it all, our shameful phallic hymn, and so dives into the night. The earth listens only to the moon.

Excised Passage from Pseudepigrapha

Wisdom is an ever-living fire, and the wise purify themselves with many flames. The fool is not want of belief, for divine law prevails inasmuch as we are fair and willing.  God is quick to convict. The fool is said to have the most beautiful of men's want for the same rivers. The science and principles of poets pass away through strife, only a soul's own walls will suffice. You cannot purify yourself with human opinions. Men feel sated by a child's knowledge. A bad witness to the one is called a beast.There is no greater sacrifice than to wash the feet of a child, and set it on a cleaner path.

Oblito Mori

This poem is much sparser but demonstrates the utilization of generative AI in a more constrained way. The final product consists of a primitive and fragmentary acrostic poem with the theme of death which I read at the funeral of my grandfather. It is no longer possible to determine the texts on which it was trained as they are no longer cataloged, though if memory serves me it was not as heavily weighted toward any specific text and was likely trained primarily on a large dictionary of English words. The programmed constraints demanded output be short, start with a given letter, and be biased to contain the same letter within the subsequent words. Much discretion was then used to reformat the output.

All buzz
Birth chord plucked
Christcrossd vibration cacophony
Death harmony
Enough ever fools fancy
Façades fragment
Moon reflect
Nowhere
Oblito mori
Pulled
Resonant bodies
Stirred shimmering static
Tears rise fall
Untuning

Maat

This final poem is also missing a definite catalog of training corpora, though I know it was trained on English translations of Egyptian morality primary sources. The title is the name of an Egyptian embodiment of the moral and ethical principles which every citizen was expected to adhere to.

Bring before your eyes the good laid up, Maat: the whole overlay of culture depends on an elaborate system for the communication of emotional bias. 

True symbols prevent infection. Coloration concealed, and with impersonal irreverence toward precepts, pursue the wind.  [but do not eat the heart]

Carefully planned desires, and the stable sameness of known and familiar acts on one's entire organism make what is high immediate; one damn thing after another. 

Comforted and strengthened, go forward. The scoffers and critics make you look good.

Conclusion

Since publishing these works I have continued to apply generative AI to the creative process in multiple media. My most time intensive project used GPT-2 trained on hundreds of translations and transliterations of the Tao Te Ching to produce another. This project made it to the front page of Hacker News via Gwern's site. The draft remains unpublished and currently consists of hundreds of pages of raw output and dozens of pages of output in various stages of editing. The landscape of generative AI writing has changed dramatically since I first experimented with it, and future publications on my part would require much stricter analysis and reference to training data. It is also apparent that some of the context window lengths may have been too large, and some of the models overfit to the training data, as there are a number of phrases which are plagiarized from corpus text. I regret not documenting my early work more rigorously as it may have been more justifiable as research than poetry. To this day I sincerely believe artificial intelligence can illuminate and inform human philosophy.